Data processing system based on ai computing power and nas storage and computer equipment

Abstract

The application relates to the fields of artificial intelligence and data storage technology, in particular to a data processing system and computer equipment based on AI computing power and NAS storage, which comprises a model state probe step: real-time collection of target computing model runtime indexes to generate demand feature fingerprints containing demand feature vectors at the current moment; a semantic value evaluation step: scanning and matching of data vector indexes based on the fingerprints to calculate semantic entropy values of to-be-processed data blocks; a storage control step: generation of an I / O priority queue according to the semantic entropy values to execute a storage scheduling strategy for the to-be-processed data blocks; and the application realizes on-demand data supply in a NAS environment, significantly improves effective semantic throughput, and accelerates the convergence process of a deep learning model.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data storage technology, specifically to a data processing system and computer equipment based on AI computing power and NAS storage. Background Technology

[0002] In existing deep learning data processing systems, the system typically includes a computing cluster and a network-attached storage device. Multiple computing nodes in the computing cluster are connected to the storage device through a high-speed network and read the dataset in the storage device through a file system protocol for model training.

[0003] Current storage control strategies mainly rely on physical layer access statistics, such as using the Least Recently Used algorithm to record the access frequency or time of data blocks. Based on this statistics, the storage controller keeps hot data in the cache and evicts cold data to maintain the responsiveness of the storage system.

[0004] However, this scheduling method based solely on physical access characteristics cannot identify the semantic relationship between data content and the upper-level computing model. This results in a large amount of redundant data that contributes little to the current model training being transmitted in the storage channel, causing a waste of computing resources and storage bandwidth, and thus leading to low model training efficiency.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a data processing system and computer equipment based on AI computing power and NAS storage. Specifically, the technical solution of this invention is as follows:

[0007] Model state probe module, semantic value assessment module, and storage control module;

[0008] The model state probe module is used to collect runtime metrics of the target computing model in real time and generate a demand feature fingerprint containing the demand feature vector at the current moment based on the runtime metrics.

[0009] The semantic value assessment module is used to receive the demand feature fingerprint, and scan and match the data vector index in the storage system based on the demand feature fingerprint to calculate the semantic entropy value of the data block to be processed. The semantic entropy value is a quantitative value that characterizes the relevance of the data block to the current model demand.

[0010] The storage control module is used to generate an I / O priority queue based on the semantic entropy value, and execute a storage scheduling strategy for the data block to be processed according to the I / O priority queue.

[0011] Preferably, the runtime metrics include any one or more of the following: the rate of change of the model loss function, the gradient norm, and the feature vector region of interest distribution.

[0012] Preferably, the data vector index is built on the metadata layer of the NAS storage system;

[0013] The data vector index contains a mapping relationship between storage logic block addresses and high-dimensional feature vectors, which is used to establish an index mapping from physical storage addresses to semantic content.

[0014] Preferably, the storage control module has a preset high-value threshold and a low-value threshold, and the high-value threshold is greater than the low-value threshold;

[0015] The storage control module is specifically used for:

[0016] In response to the semantic entropy value being greater than or equal to the high value threshold, a prefetch instruction is generated, and the data block to be processed is preloaded from the slow storage medium layer to the cache layer.

[0017] In response to the semantic entropy value being less than or equal to the low value threshold, the transmission request for the data block to be processed is intercepted, or the data block to be processed is migrated to a cold data storage area.

[0018] Preferably, the storage control module is further configured to:

[0019] In response to the semantic entropy value being between the low value threshold and the high value threshold, the current storage level of the data block to be processed is maintained, and the data block to be processed is added to the transmission queue in accordance with the standard I / O order.

[0020] Preferably, the storage control module is further configured to:

[0021] Construct a mapping function between semantic entropy values ​​and storage bandwidth allocation weights;

[0022] Using the mapping function, the target bandwidth of different data blocks in the I / O priority queue is calculated;

[0023] The transmission resource allocation of the NAS storage channel is dynamically adjusted based on the target bandwidth.

[0024] Preferably, the system further includes an in-memory computing unit;

[0025] The in-memory computing unit is integrated on the NAS storage controller side and is used to perform feature pre-screening based on the demand feature fingerprint using local computing power before the data block to be processed is output.

[0026] The storage control module is also used to calculate the semantic similarity between multiple data blocks when the system I / O is idle, and to identify multiple data blocks with semantic similarity greater than a preset reorganization threshold, and reorganize the multiple data blocks to physically contiguous storage addresses.

[0027] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the functions of a data processing system based on AI computing power and NAS storage as described in any one of the above descriptions.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. This invention achieves a deep understanding of the cognitive state of AI models by introducing multi-dimensional runtime. Compared with a single indicator, this invention uses a gradient-weighted attention aggregation algorithm to dynamically generate demand feature fingerprints and combines them with historical difficult case mining logic to accurately analyze the specific feature requirements of the model at different training stages, thus avoiding the risk of misjudgment.

[0030] 2. This invention achieves refined hierarchical management of storage resources through a preset dynamic high and low value threshold strategy in the storage control module. The system anchors the threshold in real time according to the historical distribution histogram, performs aggressive prefetching of high-entropy data to the high-speed cache, and performs interception or cold migration of low-entropy data, thereby ensuring that the valuable high-speed cache only serves critical data, effectively shielding noise interference and significantly improving the cache hit rate.

[0031] 3. This invention dynamically adjusts the transmission resource allocation of NAS storage channels by constructing an exponential mapping function between semantic entropy values ​​and storage bandwidth allocation weights, ensuring that high-value data streams can exclusively occupy high-speed channels. In multi-task concurrent scenarios, this mechanism dynamically adjusts the number of issued tokens through the QoS interface, effectively preventing low-value data streams from preempting bandwidth for critical data streams and maximizing the model accuracy gains per unit bandwidth. By introducing L2 normalization, numerical stability checks, and adaptive baseline offset update logic into the mathematical model, the robustness of semantic entropy value calculation and demand feature fingerprint generation is ensured. The system uses exponential moving average to update the baseline offset, enabling the algorithm to adapt to the dynamic drift of the feature space during training and preventing function saturation or computational failure caused by the overall shift of input values.

[0032] 4. This invention utilizes an in-memory computing unit integrated into the NAS storage controller to pre-screen data based on feature fingerprints before output, enabling the direct discarding of low-similarity invalid data payloads at the controller level. This significantly reduces host-side I / O processing load and transmission latency. By executing a semantic reorganization strategy during system I / O idle time, it identifies highly similar data blocks using locality-sensitive hashing and anchor-point linearized sorting algorithms, and reorganizes them into physically contiguous storage addresses by calling the file system interface. This scheme aggregates originally discretely distributed semantically related data in physical space, eliminates seek latency by utilizing the sequential write characteristics of the disk, and allows subsequent continuous prefetch operations to fully utilize the disk's sequential throughput bandwidth, thus solving the performance bottleneck problem of random read operations on mechanical hard drives. Attached Figure Description

[0033] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0034] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0036] Example 1:

[0037] Please see Figure 1 A data processing system based on AI computing power and NAS storage includes: a model status probe module, a semantic value assessment module, and a storage control module;

[0038] Among them, the model state probe module is used to collect the runtime indicators of the target computing model in real time, and generate a demand feature fingerprint containing the demand feature vector at the current moment based on the runtime indicators.

[0039] The semantic value assessment module is used to receive demand feature fingerprints and scan and match the data vector indexes in the storage system based on the demand feature fingerprints to calculate the semantic entropy value of the data block to be processed. The semantic entropy value is a quantitative value that characterizes the relevance of the data block to the current model demand.

[0040] The storage control module is used to generate an I / O priority queue based on the semantic entropy value, and execute the storage scheduling strategy for the data blocks to be processed according to the I / O priority queue;

[0041] This embodiment reveals in detail the core architecture and operating mechanism of the system, aiming to break the limitation of traditional storage systems that only schedule based on physical access frequency, and to build a storage-computing collaboration mechanism based on semantic value.

[0042] The system first establishes a real-time feedback link from the AI ​​computing side to the storage side through a model state probe module. This module is deployed as a plug-in in the backend interface of the deep learning framework to collect runtime metrics of the target computing model in real time. Based on these metrics, a requirement feature fingerprint is generated. The fingerprint is a high-dimensional vector whose value represents the direction of the data features that the current AI model is most confused about or most in need of learning, derived from the current gradient direction or region of interest mapping of the model.

[0043] The semantic value assessment module acts as a semantic filter, deployed on the edge computing node of the NAS storage controller, and receives... It also scans and matches the data vector indexes in the storage system's metadata layer, calculating the semantic entropy value of the data blocks to be processed during this process. To quantify the value of data blocks, this embodiment introduces a semantic entropy calculation model, the calculation formula of which is as follows:

[0044]

[0045] in, Semantic entropy value, derived from real-time calculation results, physically represents the criticality of the information contained in a data block to the correction of the current model, and is dimensionless;

[0046] The natural constant, with a value of approximately 2.71828, serves as the base of the exponential function;

[0047] : Demand feature fingerprint, which originates from the input of the model state probe module, and its physical meaning is the feature demand vector of the model at the current moment;

[0048] : Data index vector, which originates from the data feature representation pre-stored in the NAS metadata layer, and its physical meaning is the semantic content compression of the stored data;

[0049] Sensitivity coefficient, derived from system initialization settings, with a recommended value of 5.0. Its physical meaning is to adjust the response rate of the entropy value to changes in similarity. This value is based on the derivative characteristics of the Sigmoid function: when... When, the function is The derivative is maximum at the point where the input value deviates from the reference value, and when the input value deviates from the reference value. At that time, the output entropy value will generate approximately The attenuation is multiplied by a factor of 1, thereby ensuring that the system can sensitively distinguish strongly correlated data from weakly correlated data in the entropy space.

[0050] The baseline offset is derived from historical average similarity statistics and its physical meaning is the threshold used to filter background noise.

[0051] To ensure To adapt to the dynamic shifts in the feature space distribution during training, this embodiment defines specific adaptive update logic: the system maintains a global variable. And perform an exponential moving average update at the end of each training epoch:

[0052]

[0053] in, Set the smoothing factor to 0.95. For all accessed data blocks within the current Epoch and The mean cosine similarity; this formula ensures that the baseline offset always follows the current average feature response level of the model, preventing the Sigmoid function from entering the saturation region due to the overall offset of the input values;

[0054] It should be noted that, in order to ensure the numerical stability of the mathematical model, the system will perform a check before executing the above formula. and Perform L2 normalization, i.e. Ensure dot product operation The result is strictly equivalent to cosine similarity and falls within the range of Within the range; based on this, the storage control module... Generate an I / O priority queue, which is constructed as a max-heap structure. The sorting key values ​​are used to perform the final physical scheduling;

[0055] This embodiment introduces a semantic entropy-aware mechanism at the storage layer, enabling on-demand data supply in a NAS environment with limited physical bandwidth. This design allows the storage system to no longer blindly supply data, but instead prioritize responding to the high-entropy data that the model urgently needs. This improves the effective semantic throughput and accelerates the convergence process of deep learning models without increasing hardware costs.

[0056] Runtime metrics include one or more of the following: rate of change of model loss function, gradient norm, and feature vector region of interest distribution.

[0057] This embodiment further defines the specific composition and technical meaning of runtime metrics, used to characterize the starvation state of the model from multiple dimensions; the system monitors the rate of change of the model's loss function. That is, in the most recent The descent slope of the loss function value within a training step; to ensure that this metric accurately reflects the model's convergence trend and resists batch-level oscillations, this embodiment clarifies the specific calculation formula for the descent slope:

[0058] Build the latest The set of loss values ​​for each step The least squares method is used to fit the linear regression line. and the regression slope The absolute value is defined as ; defined here This is a configurable system parameter representing the size of the historical sliding window used to calculate the rate of change. Its value is typically set during system initialization to an integer between 50 and 500 based on the batch size of the training data. Larger values ​​are preferred. The value is used to smooth out the training process with high noise; a smaller value is preferred. The value is used for mutations in the fast response model;

[0059] when When the value approaches zero, it indicates that the model has entered a plateau period. At this time, the system triggers the hard example mining logic. The specific operation is: accessing the historical hard example buffer that is pre-built and maintained in local memory by the model state probe module. This buffer is used to record the metadata of the samples with the highest loss values ​​during training in real time.

[0060] The system extracts the feature vectors of the Top-K samples with the highest historical loss values ​​from the buffer and calculates their centroid vectors. The specific calculation logic is as follows: for the extracted... eigenvectors Perform L2 normalization and calculate the arithmetic mean vector. Again Normalization yields the final centroid vector. To ensure that it lies on the unit hypersphere and is directly used as the current demand feature fingerprint. This forces the storage system to retrieve data that is geometrically similar to historical difficult cases;

[0061] Building upon this, for the normal training phase outside of plateau periods, the system utilizes the gradient norm. and the distribution of feature vector regions of interest Dynamically generated using gradient-weighted attention aggregation algorithm The specific calculation logic is as follows:

[0062]

[0063] in, : The fingerprint of demand characteristics at the current moment;

[0064] The fingerprint from the previous moment. The momentum decay coefficient;

[0065] The size of the current training batch is derived from the hyperparameter settings during deep learning model training. Its physical meaning is the number of samples processed in parallel during a single iteration update, i.e., the total number of samples participating in this gradient aggregation calculation.

[0066] The current batch number The gradient L2 norm of each sample is used as the importance weight;

[0067] :No. The original feature vector of each sample;

[0068] A binary mask generated based on the attention map is used to preserve features of high-response regions;

[0069] Hadamard product, representing element-wise multiplication;

[0070] It is important to note that, in order to prevent issues arising when the model tends to converge or gradients vanish, To prevent division by zero anomalies caused by the denominator approaching zero, the system performs a numerical stability check before performing the division operation: if the denominator... Less than the preset minimum floating-point value ,For example If the system determines that the current batch cannot provide effective directional guidance, it will directly set the batch to... This means maintaining the demand fingerprint from the previous time step unchanged, thus ensuring the robustness of the algorithm; furthermore, to strictly satisfy the geometric assumptions of the semantic value evaluation module regarding the unit hypersphere of the input vector, the system calculates... Then, L2 normalization will be performed: Ensure that the output fingerprint has a modulus of 1.

[0071] To ensure consistency in the operational dimensions of the Hadamard product in the above formula and to eliminate the black-box properties based on attention map descriptions, this embodiment explicitly stipulates: This is the channel feature vector output from the penultimate layer of the model, with dimension . ; The specific generation algorithm is as follows: calculate the loss function of the current batch. Regarding eigenvectors gradient vector ,right The vector is obtained by taking the absolute value of each element. And based on the formula Robust normalization is obtained ,in Used to prevent the denominator from being zero, if detected If the gradient is flat, then directly... Set all values ​​to 1 to preserve all features;

[0072] Set dynamic mask threshold for The average of all elements in the set, if The Middle The value of each element is greater than Then the corresponding Set to 1 if necessary, otherwise set to 0; this process ensures This formula can accurately filter out feature channels that do not contribute to the current gradient descent direction; it ensures that... It can track the factors that contribute the most to the model parameter updates in real time. Large, and the model is most concerned with, namely The coverage of feature directions avoids the risk of misjudgment caused by a single indicator;

[0073] This embodiment achieves a deep understanding of the cognitive state of the AI ​​model by introducing multi-dimensional runtime metrics. This mechanism can not only determine whether the model is learning, but also analyze what features the model wants to learn, thereby avoiding the risk of misjudgment caused by a single metric and ensuring the accuracy and robustness of the generation of demand feature fingerprints.

[0074] The data vector index is built on the metadata layer of the NAS storage system; it contains the mapping relationship between storage logical block addresses and high-dimensional feature vectors, and is used to establish an index mapping from physical storage addresses to semantic content.

[0075] This embodiment details the construction method of the data vector index, aiming to solve the problem of the lack of semantic information in traditional NAS metadata; the system constructs a heterogeneous extended index at the metadata layer of the NAS storage system, and this data vector index contains the following mapping relationships:

[0076]

[0077] in, : The address of the logical block is stored, which comes from the file system allocation table and has the physical meaning of the physical location of the data on the hard disk;

[0078] High-dimensional feature vectors originate from preprocessing calculations during data writing and physically represent the semantically compressed representation of the data block content; the key is to ensure... The probe module generated Being in the same vector space gives mathematical meaning to the dot product operation, generating The preprocessing algorithm is configured to share Backbone weights with the target computation model or employ a student network aligned with knowledge distillation;

[0079] For student networks aligned using knowledge distillation, this embodiment clarifies the corresponding alignment constraints: the system deploys a lightweight network, such as MobileNet, as the Student model during the pre-training phase and constructs a hybrid loss function. ,in Preset weights; by minimizing this loss function, the vectors generated on the storage side are forced to... It maintains strict collinearity with the high-dimensional feature space of the computational model in the geometric direction, eliminating metric failure caused by differences in model structure;

[0080] To ensure the long-term consistency of the vector space and prevent issues arising from parameter updates during the training of the target computational model. and To address spatial mismatch, the preprocessing algorithm employs either a frozen backbone strategy or a spatial mapping strategy; whereby the frozen backbone strategy refers to the strategy used to generate... The network weights are fixed to the pre-trained weights before model training; the spatial mapping strategy refers to the system maintaining a linear mapping matrix. The dynamics generated by the current model Project to In the static feature space, that is ;

[0081] In order to solve To address the parameter acquisition issue, this embodiment employs an online calibration mechanism based on orthogonal Prouke analysis: the system performs calibration every... Number of training steps, such as Sample a set of current dynamic features and corresponding static index features Construct the covariance matrix For the matrix Performing singular value decomposition yields and order ;

[0082] This calculation process guarantees that, in the least squares sense... Minimize to accurately align the drifted feature space; for a specific implementation of sampling a set, to ensure strict pairing of input data, the system employs a batch metadata backtracking mechanism: after the training steps meet the requirements... At that time, the system records the file inode number of each sample in the current batch and its corresponding model output features;

[0083] Through the InodeID in the B+ tree of the NAS metadata layer The complexity accurately retrieves the corresponding original storage vector. This allows for the construction of strictly paired calibration datasets. This eliminates noise introduced by data misalignment; in addition, SVD decomposition calculations are configured to be asynchronously distributed to the management node of the AI ​​cluster for execution, avoiding the occupation of the NAS controller's real-time I / O scheduling computing power.

[0084] The construction process is as follows: When new data is written to the NAS, the system automatically extracts data features using idle computing power to generate... and with The data is bound to a B+ tree or HNSW index structure in the metadata area; this design enables a direct index mapping from physical storage address to semantic content, allowing the storage system to understand the meaning of the data without fully reading it.

[0085] This embodiment achieves nanosecond-level semantic value assessment by building a heterogeneous extended index at the metadata layer. This technical solution can complete the value judgment by scanning the metadata without touching the underlying data entities, reducing the I / O latency of semantic awareness and making it possible to filter large-scale datasets in real time.

[0086] The storage control module has a preset high value threshold and a low value threshold, and the high value threshold is greater than the low value threshold. Specifically, the storage control module is used to: generate a prefetch instruction in response to a semantic entropy value greater than or equal to the high value threshold, and preload the data block to be processed from the slow storage medium layer to the cache layer; and intercept the transmission request of the data block to be processed or migrate the data block to be processed to the cold data storage area in response to a semantic entropy value less than or equal to the low value threshold.

[0087] This embodiment details the threshold-based decision-making logic of the storage control module, aiming to achieve refined hierarchical management of storage resources; the storage control module has preset high-value thresholds. and low value threshold And satisfy To ensure the robustness of the threshold setting and adapt to the entropy distribution drift at different training stages, this embodiment abandons the fixed value setting and instead adopts a dynamic statistical strategy based on historical sliding windows.

[0088] Specifically, the system maintains a list covering the most recent minutes, for example A histogram showing the historical distribution of semantic entropy values ​​for all processed data blocks within the data block; The real-time anchor is the 80th percentile value of this histogram. Anchored to the 20th percentile value; this mechanism can effectively smooth out noise caused by instantaneous queue jitter or data differences between batches, ensuring that the grading strategy is always based on the macro value distribution.

[0089] The specific execution logic is as follows: First, respond to the semantic entropy value. The system triggers an aggressive prefetch strategy, generating prefetch instructions and bypassing the standard file system prefetch algorithm. It directly loads the data block corresponding to the logical block address from the slow storage medium layer to the cache layer in parallel, ensuring zero computational wait time. In response to... The system triggers an interception and cold migration strategy, directly canceling the data block's position in the transmission request queue or placing it at the end of the queue, and marking the data block as semantically cold data in the background, thereby migrating it to the cold data storage area;

[0090] This embodiment significantly optimizes the cache hit rate and bandwidth utilization of the storage system through a polarized threshold control strategy. This mechanism ensures that the valuable high-speed cache serves only the data most critical to model correction, while effectively shielding the noise interference of low-value data and solving the I / O bottleneck problem in massive data training scenarios.

[0091] The storage control module is also used to: maintain the current storage level of the data block to be processed in response to the semantic entropy value being between the low value threshold and the high value threshold, and add the data block to be processed to the transmission queue in accordance with the standard I / O order;

[0092] This embodiment details the processing method for intermediate state data between high and low thresholds, reflecting the system's stability control concept; specifically, it responds to semantic entropy values. satisfy The system executes a storage tier maintenance strategy, without performing active data migration or prefetching operations, and maintains the data block in its current physical medium location; the system adds the data block to be processed to the transmission queue for processing according to the standard I / O order, that is, the FIFO order based on the request arrival time or the elevator scheduling algorithm based on the physical address continuity;

[0093] This embodiment employs a robust scheduling strategy for medium-value data, avoiding system jitter and additional metadata overhead caused by frequent cross-level data movement. This design maintains the baseline I / O performance of the system while ensuring the priority of high-value data, thus ensuring the smoothness of the overall data pipeline.

[0094] The storage control module is also used to: construct a mapping function between semantic entropy values ​​and storage bandwidth allocation weights;

[0095] Calculate the target bandwidth for different data blocks in the I / O priority queue using a mapping function;

[0096] Dynamically adjust the transmission resource allocation of the NAS storage channel based on the target bandwidth;

[0097] This embodiment details a dynamic bandwidth allocation mechanism based on semantic entropy values, aiming to maximize the model accuracy gains per unit bandwidth. The storage control module constructs a bandwidth mapping function to calculate the target bandwidth for different data blocks in the I / O priority queue. The bandwidth allocation model is as follows:

[0098]

[0099] in, Target bandwidth, derived from function calculation output, physically represents the transmission rate quota allocated to the current data stream, measured in MB / s;

[0100] The total bandwidth of the NAS storage channel originates from hardware physical attributes, and its physical meaning refers to the physical upper limit of the network interface card.

[0101] The semantic entropy value of the current data block is derived from the semantic value assessment module.

[0102] The first person involved in bandwidth allocation calculation The semantic entropy value of each data block is derived from the system's real-time entropy statistics of the current concurrent data stream;

[0103] The number of concurrent data streams currently being transmitted is based on real-time system monitoring.

[0104] The weight adjustment factor is a preset constant, and its physical meaning is to control the sensitivity of bandwidth allocation differentiation.

[0105] Based on this, the storage control module calculates... By calling the QoS interface of the underlying storage driver, the number of transmit tokens for each I / O queue is dynamically adjusted; the specific conversion logic is as follows: to prevent [further issues caused by QoS issues]... If the value is too small, the integer will be truncated to zero, causing low-priority queues to starve. The system uses a unit conversion formula with a minimum guarantee: Let the unit size of the token bucket be... If it is 4KB, then the token filling rate is... Here it is assumed The unit is MB / s; this logic ensures that any active queue has at least the minimum bandwidth to maintain a connection heartbeat; the system will calculate... Write it into the flow control register of the network interface card controller to achieve physical layer resource reallocation;

[0106] To prevent control plane oscillations caused by frequent overwriting of the flow control register due to instantaneous entropy fluctuations, this embodiment introduces a control cycle latching mechanism: setting a global bandwidth update cycle. For example, 50ms; in Within the window, the system only accumulates the entropy statistics for each queue. And at the end of the window, a division and register write operation are performed uniformly; this mechanism limits the control overhead to less than 1% of the total CPU cycles, while ensuring the smoothness of bandwidth allocation.

[0107] This embodiment breaks away from the traditional egalitarian bandwidth allocation model of storage systems. Through exponential weight mapping, it ensures that high-value data can exclusively occupy high-speed channels. In multi-task concurrent scenarios, this mechanism effectively prevents low-value data streams from preempting the bandwidth of critical data streams, thereby improving the training efficiency of AI clusters on a macro level.

[0108] The system also includes an in-memory computing unit; the in-memory computing unit is integrated on the NAS storage controller side and is used to perform feature pre-screening based on demand feature fingerprints using local computing power before outputting the data blocks to be processed; the storage control module is also used to calculate the semantic similarity between multiple data blocks when the system I / O is idle, and to identify multiple data blocks with semantic similarity greater than a preset reorganization threshold, and reorganize the multiple data blocks to physically contiguous storage addresses.

[0109] This embodiment details the implementation of in-memory computing and physical reassembly technologies, aiming to further reduce host load and optimize physical read performance. The system introduces an in-memory computing unit integrated on the NAS storage controller SoC or dedicated FPGA card. When a data block leaves the disk medium and passes through the controller, this unit uses local computing power to directly compare the feature vector of the data block with the required feature fingerprint. The preset coarse screening threshold is set to This value originates from the statistical analysis of the similarity distribution of a large number of unrelated sample pairs during the system initialization phase, selecting the [number]th [sample]. The percentile value is used as a cutoff line, and its theoretical basis is the principle of orthogonality in high-dimensional spaces: in dimension... In the high-dimensional feature space, any two random vectors are approximately orthogonal, and the mean of the cosine similarity distribution tends to 0;

[0110] According to the theory of concentration inequalities, the probability of similarity exceeding 0.3 is extremely low. Therefore, this threshold can effectively define the statistical boundary between semantic relevance and random noise, aiming to filter out the vast majority of background noise with low computational cost. If the similarity is lower than the preset coarse screening threshold, which is set to 0.3, this value comes from the statistical analysis of the similarity distribution of a large number of unrelated sample pairs during the system initialization phase. The 95th percentile value is selected as the cutoff line. If the similarity is lower than the preset coarse screening threshold, which is set to 0.3, this threshold can be used to filter out the vast majority of background noise with low computational cost. In this case, the transmission load of the data block is discarded directly at the controller level, and only an invalid label is returned.

[0111] During this process, the in-memory computing unit does not directly read the complete data block content to recalculate the features, but instead obtains the corresponding feature vector by directly accessing the data vector index in the controller cache. This minimizes computational latency. In addition, when the storage control module detects that the system I / O is idle, it executes a semantic reorganization strategy, calculates the mutual semantic similarity between multiple data blocks in the storage area, and identifies a set of data blocks with a mutual similarity greater than a preset reorganization threshold.

[0112] To balance computational efficiency and coverage, the system does not perform pairwise comparisons of all stored data in the step of calculating semantic similarity between multiple data blocks. Instead, it employs a sliding window mechanism based on access frequency, selecting only the data within the most recent time window. The frequency of internal access exceeds the threshold The set of active data blocks is used as the computation object;

[0113] Among them, parameters The value of is set to 1.5 times the average time required to complete one epoch of the current training task, for example, 30 minutes, to cover the most recent data iteration cycle; Parameter The value is set to the average number of accesses to all accessed data blocks within the time window, used to dynamically define the active boundary; simultaneously, the Locality Sensitive Hashing (LSH) algorithm is used to quickly filter candidate similar pairs, performing precise cosine similarity calculations only on data block pairs that collide with LSH, thereby reducing computational complexity from... Reduced to approximately ;

[0114] To meet the code-level reproducibility requirements, the locality-sensitive hashing algorithm is specifically configured as a random hyperplane projection method suitable for cosine similarity measurement; the system pre-initializes a random projection matrix. ,in For feature vectors Dimensions The hash fingerprint has a bit length, for example 64, and the matrix elements follow a standard normal distribution. To ensure the consistency of hash fingerprints generated by each node in the distributed storage cluster, the system uses the Mason tween algorithm as a pseudo-random number generator and sets a globally uniform random seed, for example, Seed=12345. It then uses Box-Muller transformation to generate matrix parameters that strictly satisfy a normal distribution, thus guaranteeing consistency. Deterministic reconstruction of a matrix across different computation nodes;

[0115] For any data block's feature vector Calculate its binary hash fingerprint ,in The sign function takes the value 1 for positive numbers and 0 for negative numbers. The system only marks the data blocks corresponding to fingerprint pairs with a Hamming distance of less than or equal to 3 as candidate similar pairs, and then performs precise cosine similarity calculation to determine whether the recombination threshold is exceeded.

[0116] Regarding the correspondence between the recombination threshold of 0.85 and the LSH parameter, this embodiment is based on the geometric properties of SimHash. Perform calibration; when similarity At that time, single-bit collision probability To ensure a high recall rate, the system sets a maximum allowed Hamming distance. This parameter combination, verified by Monte Carlo simulation, can effectively cover more than 95% of highly similar samples, avoiding missed detections.

[0117] For reorganized to physically contiguous storage addresses, the specific execution process is as follows: the storage controller calls the file system interface, such as fallocate, to allocate a physically contiguous free extent on the outer ring of the disk;

[0118] Execute the anchor point linearization sorting algorithm: In order to transform the discrete LSH query results into a sortable set of entities, the system first performs a semantic clustering operation: selects the block with the highest access frequency in the active data block set as the semantic anchor point, and retrieves all Hamming distances to its LSH fingerprint. The data blocks form a candidate cluster for reorganization; the data block with the highest access frequency in this set is selected as the anchor point. Popularity of visits here Defined using a time decay cumulative model:

[0119]

[0120] in, For the data block within the window Next visit time. Take the attenuation coefficient as This definition ensures that the most recently accessed data block has a higher sorting weight than simply counting access frequency; the remaining data blocks in the set are calculated. and cosine similarity and according to Arrange the candidates in descending order to generate the initial candidate sequence;

[0121] The system performs semantic truncation, traversing the initial candidate sequence and extracting the relevant segments. For consecutive segments that are consistently greater than the preset recombination threshold of 0.85, long-tail data that do not meet the similarity requirements are removed, thereby constructing a recombination target set that strictly satisfies the high semantic relevance; this serves as a concrete implementation of the clustering order based on similarity, ensuring that physical adjacency and semantic relevance are positively correlated and monotonically mapped.

[0122] The buffer data in the recombined target set is written into the newly allocated Extent in the above sequence order, and the seek latency is eliminated by using the sequential write feature.

[0123] The logical-physical mapping pointer in the metadata tree is atomically updated, pointing the original logical address to the new physical extent and marking the original physical block as pending reclamation. This process physically aggregates the originally discrete semantically related data, enabling subsequent continuous prefetch operations to fully utilize the sequential throughput bandwidth of the disk, thus solving the problem of the random read performance bottleneck of mechanical hard drives.

[0124] To address the issue of disk fragmentation preventing the allocation of large contiguous blocks of space at once, the system introduces a fallback reorganization strategy: if fallocate returns an ENOSPC or EOPNOTSUPP error, the system will split the reorganization target set into two parts and attempt to request two smaller extents; if the size of the split request is less than 4MB, which is the minimum sequential read benefit threshold set by the system.

[0125] The reassembly operation for that subset is then terminated to avoid generating new fragments, thereby ensuring the system's robustness on aging file systems; this threshold The settings are based on the disk. Performance model: To reduce the transmission time of data blocks Significantly greater than the average seek time ,set up ;by and transmission rate For example, the calculation is as follows The value is rounded down to 4MB, thus ensuring the throughput advantage of sequential reads.

[0126] Example 2:

[0127] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the functions of a data processing system based on AI computing power and NAS storage.

[0128] This embodiment provides a specific hardware implementation scheme for a computer device, which includes a processor, a memory, and a computer program; wherein, the processor may be a central processing unit (CPU), a graphics processing unit (GPU), a dedicated AI acceleration chip (NPU), or a field-programmable gate array (FPGA); the memory includes volatile memory and non-volatile memory, used to store computer program instructions;

[0129] When the processor executes a computer program, it specifically implements any of the above functions, including but not limited to: collecting runtime metrics and generating demand feature fingerprints, calculating semantic entropy values ​​based on fingerprints, generating priority queues based on entropy values ​​and performing memory scheduling, and performing data prefetching, interception, bandwidth allocation, and physical reassembly.

[0130] This embodiment provides a practically deployable physical entity by embedding the above-mentioned semantic value-based in-memory computing collaboration algorithm into a specific hardware device. This device enables the technical solution of the present invention to be applied in real-world scenarios such as data centers and edge computing nodes, providing hardware support for building high-performance AI infrastructure.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data processing system based on AI computing power and NAS storage, characterized in that, It includes a model state probe module, a semantic value assessment module, and a storage control module; The model state probe module is used to collect runtime metrics of the target computing model in real time and generate a demand feature fingerprint containing the demand feature vector at the current moment based on the runtime metrics. The semantic value assessment module is used to receive the demand feature fingerprint, and scan and match the data vector index in the storage system based on the demand feature fingerprint to calculate the semantic entropy value of the data block to be processed. The semantic entropy value is a quantitative value that characterizes the relevance of the data block to the current model demand. The storage control module is used to generate an I / O priority queue based on the semantic entropy value, and execute a storage scheduling strategy for the data block to be processed according to the I / O priority queue. 2.The AI algorithm power and NAS storage-based data processing system according to claim 1, wherein, The runtime metrics include any one or more of the following: the rate of change of the model loss function, the gradient norm, and the feature vector interest region distribution. 3.The AI algorithm power and NAS storage-based data processing system of claim 1, wherein, The data vector index is built into the metadata layer of the NAS storage system; The data vector index contains a mapping relationship between storage logic block addresses and high-dimensional feature vectors, which is used to establish an index mapping from physical storage addresses to semantic content. 4.The AI algorithm power and NAS storage-based data processing system of claim 1, wherein, The storage control module has a preset high value threshold and a low value threshold, and the high value threshold is greater than the low value threshold. The storage control module is specifically used for: In response to the semantic entropy value being greater than or equal to the high value threshold, a prefetch instruction is generated, and the data block to be processed is preloaded from the slow storage medium layer to the cache layer. In response to the semantic entropy value being less than or equal to the low value threshold, the transmission request for the data block to be processed is intercepted, or the data block to be processed is migrated to a cold data storage area. 5.The AI algorithm power and NAS storage-based data processing system according to claim 4, wherein, The storage control module is also used for: In response to the semantic entropy value being between the low value threshold and the high value threshold, the current storage level of the data block to be processed is maintained, and the data block to be processed is added to the transmission queue in accordance with the standard I / O order. 6.The AI algorithm power and NAS storage-based data processing system of claim 1, wherein, The storage control module is also used for: Construct a mapping function between semantic entropy values ​​and storage bandwidth allocation weights; Using the mapping function, the target bandwidth of different data blocks in the I / O priority queue is calculated; The transmission resource allocation of the NAS storage channel is dynamically adjusted based on the target bandwidth.

7. The AI algorithm power and NAS storage-based data processing system according to any one of claims 1 to 6, characterized in that, The system also includes an in-memory computing unit; The in-memory computing unit is integrated on the NAS storage controller side and is used to perform feature pre-screening based on the demand feature fingerprint using local computing power before the data block to be processed is output. The storage control module is also used to calculate the semantic similarity between multiple data blocks when the system I / O is idle, and to identify multiple data blocks with semantic similarity greater than a preset reorganization threshold, and reorganize the multiple data blocks to physically contiguous storage addresses.

8. A computer device, comprising: It includes a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to perform the functions of the data processing system based on AI computing power and NAS storage as described in any of the preceding claims.

Images